Research map / Detection and segmentation / CNN features
Object Detection Networks on Convolutional Feature Maps
Shaoqing Ren, Kaiming He, Ross Girshick, Xiangyu Zhang, Jian Zhong Sun
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace SVM on RoI features with a 4fc MLP NoC (ablation) | NoC on ZF · PASCAL VOC 07 test | mAP (%): 45.8 -> 53.6 | Improved |
| increase MLP NoC depth from 2 fc to 4 fc layers (ablation) | NoC on ZF · PASCAL VOC 07 test | mAP (%): 49.0 -> 53.1 -> 53.6 | Improved |
| add 2 conv layers to the 3fc NoC (07+12 training) (ablation) | NoC on ZF · PASCAL VOC 07 test | mAP (%): 56.5 -> 58.9 | Improved |
| add 2 conv layers to the 3fc NoC (07 trainval only) (ablation) | NoC on ZF · PASCAL VOC 07 test | mAP (%): 53.1 -> 51.4 | Hurt |
| add 3rd conv layer beyond 2 conv (ablation) | NoC on ZF · PASCAL VOC 07 test (07+12) | mAP (%): 58.9 -> 58.8 | No clear effect |
| add maxout for scale selection after conv+1 (ablation) | 2conv3fc NoC on ZF · PASCAL VOC 07 test (07+12) | mAP (%): 58.9 -> 60.7 | Improved |
| use pre-trained fc init instead of random init (3fc NoC, 07+12) (ablation) | NoC on ZF · PASCAL VOC 07 test | mAP (%): 56.5 -> 58.0 | Improved |
| fine-tune maxout 2conv3fc NoC from pre-trained instead of random init (ablation) | NoC on ZF · PASCAL VOC 07 test (07+12) | mAP (%): 60.7 -> 62.9 | Improved |
| use deeper shared features (VGG-16 to VGG-19) (ablation) | 3fc NoC · PASCAL VOC 07 test | mAP (%): 64.6 -> 65.1 | Improved |
| increase classifier depth (3fc to 2conv3fc) on VGG-16 features (ablation) | NoC · PASCAL VOC 07 test | mAP (%): 64.6 -> 66.1 | Improved |
| add maxout 2conv3fc NoC to reduce localization error (ablation) | NoC · PASCAL VOC 07 (error analysis) | localization error (%): 28.1 -> 22.6 (ZF), 24.8 -> 20.1 (VGG-16) | Improved |
| add 3fc MLP NoC to naive Faster R-CNN (ablation) | Faster R-CNN ResNet-101 · MS COCO val | AP (%): 21.3 -> 26.3 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 0 of 1 on most shared results.
- Faster R-CNN (07+12) — Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016) · better on 0 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- R-FCN: Object Detection via Region-based Fully Convolutional Networks (NeurIPS 2016 · similarity 0.92)
- Region-Based Convolutional Networks for Accurate Object Detection and Segmentation (TPAMI 2016 · similarity 0.92)
- Feature Pyramid Networks for Object Detection (CVPR 2017 · similarity 0.92)
- Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks (TPAMI 2016 · similarity 0.90)
- Spatially Invariant Unsupervised Object Detection with Convolutional Neural Networks (AAAI 2019 · similarity 0.89)
- Deep Learning for Generic Object Detection: A Survey (IJCV 2019 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
